A Short-Term Residential Electricity Consumption Prediction Method Based on Feature Fusion Graph Neural Network

Abstract

The present invention relates to a short-term residential electricity consumption prediction method based on a feature fusion graph neural network, belonging to the technical field of short-term electricity consumption prediction in power systems. It solves the deficiencies of traditional methods in capturing spatio-temporal dependence relationships and feature fusion. The main technical points include: collecting and preprocessing publicly available real-world residential historical electricity consumption data and auxiliary feature data, and processing them separately through the input model; building a parallel time convolution structure; constructing a three-dimensional dynamic adjacency matrix through reverse tri-modal factor analysis; designing a multi-feature fusion module to dynamically introduce the spatio-temporal relationship of auxiliary features into the main feature prediction; introducing residual connection and skip connection structures, and finally constructing a multi-feature fusion spatio-temporal graph neural network model, and verifying the effectiveness of the model through evaluation indicators. The results show that the present invention can improve the accuracy of power dispatching and energy consumption management, and provide technical support for smart grid and renewable energy management.

Description

Technical Field

[0001] The present invention relates to the technical field of short-term electricity consumption prediction in power systems, and particularly to a short-term residential electricity consumption prediction method based on feature fusion graph neural network. Background Art

[0002] With the rapid growth of electricity demand in modern society, short-term residential electricity consumption prediction has become a key link in power system management. Accurate electricity consumption prediction is of great significance for power grid dispatching optimization, energy consumption management, and efficient utilization of renewable energy. However, due to the influence of various complex factors such as weather, time, and user behavior on residential electricity consumption patterns, it has significant nonlinearity and spatio-temporal dependence, and traditional prediction methods are difficult to effectively capture these characteristics.

[0003] Currently, time series-based prediction methods mainly focus on the temporal dimension changes of historical data, ignoring the potential spatial correlations between residences. In addition, single-feature-based prediction models usually cannot fully utilize multi-source data, such as auxiliary features like weather, holidays, and user behavior, which limits the prediction accuracy of the models.

[0004] In recent years, graph neural network (GNN) has received extensive attention due to its ability to process non-Euclidean data structures and has shown significant advantages in modeling complex relationships between nodes. However, existing graph neural network frameworks still have deficiencies in combining spatio-temporal characteristics and multi-feature fusion, and it is difficult to balance prediction accuracy and model robustness simultaneously. Summary of the Invention

[0005] A short-term residential electricity consumption prediction method based on feature fusion graph neural network, characterized in that the method includes the following operations:

[0006] First, use the public dataset provided by the OpenEI official website, mainly select the historical electricity consumption data of 15 residences in Louisiana (LA), USA, and the historical data of 2 auxiliary features of the residences.

[0007] Second, process the residential electricity consumption dataset and the auxiliary feature dataset respectively through the input module and compile them into a set of datasets.

[0008] Third, build a parallel time convolution structure to represent the time relationship between electricity consumptions, introduce a three-dimensional dynamic adjacency matrix to represent the complex spatial relationship between electricity consumptions, and construct a dynamic graph convolution structure.

[0009] Then, design a multi-feature fusion structure to dynamically add auxiliary features to the prediction of electricity consumption.

[0010] Finally, the entire multi-feature fusion spatio-temporal graph neural network is constructed, and the subsequent entire model is represented by DFF-GWN. The model includes input and output modules, a time module, a space module, and a multi-feature fusion module. Then, a public dataset is used to verify the model, and evaluation metrics are used to judge the effectiveness of the model.

[0011] The present invention is a method for predicting short-term residential electricity consumption using a feature fusion graph neural network. There are many auxiliary features in the dataset used. For the selection of these auxiliary features, the correlation coefficient analysis method is used to calculate the coefficient between each auxiliary feature and electricity consumption, and the two selected auxiliary features are determined according to the coefficient size. The correlation coefficient comprehensively considers two types, the Pearson coefficient and the Spearman coefficient. The formulas for the Pearson coefficient and the Spearman coefficient are as follows:

[0012]

[0013] In the formula, r and ρ are the Pearson coefficient and the Spearman coefficient respectively, x i and y i are the observed values of the auxiliary feature and electricity consumption respectively, and are the means of the auxiliary feature and electricity consumption respectively, d i is the ranking difference between x i and y i ; T is the time step; and for the electricity consumption data, the sum of all data is used as the main feature, that is, the prediction object.

[0014] For the processing of electricity consumption and auxiliary features, data normalization is performed separately, and then all data are concatenated in the feature dimension, rather than directly preprocessing all data and combining them into a single dataset, which is convenient for inputting into the model separately.

[0015] The steps for constructing the three-dimensional dynamic adjacency matrix proposed by the present invention include:

[0016] Divide the total time step T into multiple cycles by day, and the number of time slots in each cycle is N t , and then construct an adjacency matrix for each cycle.

[0017] Using the reverse tri-modal factor analysis method, reconstruct the original adjacency matrix by constructing 3 factor loading matrices and 1 core matrix Incorporate the time relationship into the construction of the adjacency matrix.

[0018] During this process, a total of 3 types of adjacency matrices are generated, namely the adjacency matrix representing the main feature the adjacency matrix representing the auxiliary feature Here a represents only one auxiliary feature, and the other feature is represented by b. In the subsequent analysis, only c is used to replace all auxiliary features to facilitate subsequent analysis. It represents the adjacency matrix of the fusion of the main feature and the auxiliary feature. N m Represents the number of main feature nodes, N c Represents the number of auxiliary feature nodes.

[0019] The three 3D dynamic graph structures generated are: main feature graph Auxiliary feature map And the main feature and auxiliary feature fusion map

[0020] The construction process of the three-dimensional dynamic adjacency matrix specifically includes:

[0021] Step (1) utilizes the temporal correlation of residential electricity consumption in adjacent time periods to share a graph structure for the 24-hour electricity consumption relationship, thereby dividing all time steps into 24-hour cycles, where N t =24, thus converting the original adjacency matrix The complexity is reduced to

[0022] The goal of the trimodal factor analysis method used in step (2) is to construct a core matrix G m，p，q And the factor loading matrix A t，m , B i，p , C j，q To reconstruct the original data matrix A t，i，j , and the multiple matrices constructed are low-dimensional matrices, thus achieving the purpose of dimensionality reduction. The core formula is as follows:

[0023]

[0024] In the formula, is the slot factor loading matrix, is the input node factor loading matrix, is the output node factor loading matrix and is the core matrix, where N i is the number of input nodes, N o is the number of output nodes, d is the embedding dimension, and the factor loading matrix A here t，m , B i，p , C j，q is a two-dimensional matrix, G m，p，q is a three-dimensional matrix,

[0025]

[0026] The above two formulas are respectively the non - negativity processing and row normalization processing of the constructed adjacency matrix, resulting in the final three - dimensional dynamic adjacency matrix used.

[0027] Step (3) is to construct three three - dimensional adjacency matrices that meet the model's requirements according to the three - dimensional adjacency matrix construction method in step (2). They are the three - dimensional adjacency matrices of the main feature and the auxiliary features respectively and Here, since multiple auxiliary features are used, only one is used to represent all auxiliary features for convenience of description. And is the three - dimensional adjacency matrix used in the multi - feature fusion module. Different from the conventional method of directly splicing the auxiliary features to the main feature for prediction, in the multi - feature fusion module, by constructing to dynamically simulate the spatio - temporal relationship between the main feature and the auxiliary features.

[0028] Step (4) is to generate a graph structure based on the three three - dimensional adjacency matrices obtained in step (3). A graph g=(v, e) consists of a set of node sets v and a set of edge sets e. Among them, v i , v j ∈V represents a node, and ε ij ∈e represents the edge between v i and v j . And these three graph structures are exactly the connections between the electricity consumption of multi - household residences and their auxiliary features, representing their spatial dependence relationship. In this process, a diffusion graph convolutional network is used to handle the spatial dependence between residences by aggregating the hidden states of adjacent residences. The formula is:

[0029]

[0030] In the formula, is the input of the l th dynamic graph convolutional layer, and O l is the output after being updated by the dynamic graph convolutional layer. Similarly, the output of the main feature is The output of the auxiliary feature is W k is the learnable parameter for K th diffusion. K is the total number of diffusion steps, thus generating the main feature graph g m , the auxiliary feature graph g c and the fused graph g cm of the main feature and the auxiliary feature.

[0031] The multi-feature fusion structure proposed by the present invention dynamically adds auxiliary features to the prediction of electricity consumption. Specifically, in the prediction of residential electricity consumption, at the same moment, the auxiliary features will not only affect the main features at the current moment, but also have an impact on the main features at adjacent moments. This relationship can be understood as a progressive relationship in which the auxiliary features affect the main features through their mutual influence. And this module takes into account the spatio-temporal dependence relationship existing between the nodes of the auxiliary features. In each layer, a separate spatio-temporal block is assigned to the main features and the auxiliary features, and the outputs of the main features and the auxiliary features are obtained, which are respectively input into their respective multi-feature fusion modules, that is, g cm and g m , and after an aggregation process, the finally fused output is obtained. The formula is:

[0032] In the formula, DGC(·) is the dynamic graph convolutional layer, and are the outputs of the main features and the auxiliary features after passing through the dynamic graph convolutional layer, is the output after multi-feature fusion. Agg(·) represents the aggregation process of the two results. The purpose of using direct summation is to make the calculation of the entire model concise and maintain a high representation ability. The adjacency matrix at time slot t is expressed as where ω(t) is a function to obtain the number of time slots in a period, and W represents the learnable parameter, represents the output result after passing through the multi-feature fusion module.

[0033] The entire multi-feature fusion spatio-temporal graph neural network model constructed by the present invention includes an input and output module, which transforms the dimension of the data through a two-dimensional convolutional network and inputs it into the model, and outputs the prediction result after passing the data through the spatio-temporal module through an activation function and a two-dimensional convolutional network. In order to improve the performance and training efficiency of the model in the entire network, a residual connection and a skip connection structure are also added. The skip connection structure integrates the spatio-temporal dependence relationships at different time scales. The formula is:

[0034] In the formula, concat(·) is the concatenation operation, is the output result of the main features of the time module in each layer, and L is the number of spatio-temporal layers.

[0035] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: (1) The present invention processes auxiliary features differently from the previous direct splicing. Instead, it processes them separately and allocates independent spatio-temporal modules to capture the spatio-temporal correlation of auxiliary features; (2) The present invention uses a diffusion convolutional layer to dynamically update the three-dimensional adjacency matrix and can operate on the dynamic graphs corresponding to different moments; (3) The present invention uses a three-modal factor method to reconstruct the three-dimensional adjacency matrix, incorporating temporal correlation into the construction of the adjacency matrix, reducing the complexity of adjacency matrix modeling, and more delicately capturing the spatial dependence relationship between residential electricity consumption; (4) The present invention designs a multi-feature fusion module that uses the three-dimensional adjacency matrix to dynamically aggregate the spatio-temporal relationships of main features and auxiliary features, improving the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a specific flowchart of a short-term residential electricity consumption prediction method based on a feature fusion graph neural network according to the present invention;

[0037] Figure 2 is a structural diagram of a DFF-GWN short-term residential electricity consumption prediction model in an example of the present invention;

[0038] Figure 3 is a diagram of the construction of a three-dimensional dynamic adjacency matrix in an example of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0039] The present invention will be further described below with reference to the accompanying drawings. The embodiments are for more clearly illustrating the technical solutions of the present invention. The scope of use of the present invention is not limited thereto and can be applied to multiple fields. Here, only an example is given.

[0040] As Figure 1 shown, an example of the present invention provides a short-term residential electricity consumption prediction method based on a feature fusion graph neural network, which is specifically divided into the following steps:

[0041] First, use the public dataset provided by the OpenEI official website, mainly select the historical electricity consumption data of 15 residences in Louisiana (LA), USA, and the historical data of 2 auxiliary features of the residences. Among them, the historical electricity consumption data of these 15 residences includes the historical electricity consumption data per hour in 2012. Each residence records the electricity consumption of various electrical equipment per hour, and a total of 8760 data points are recorded. The 2 auxiliary features are Electricity:Facility (EF) and Electricity:HVAC (EH).

[0042] Second, process the residential electricity consumption dataset and the 2 auxiliary feature datasets respectively through the input module and compile them into a set of datasets. Assume that the existing residential electricity consumption dataset and the 2 auxiliary feature datasets are and Collectively referred to as X 1：T , where 1:T represents that the historical time step is T, and only will be used to represent 2 auxiliary features in the following example analysis. The input module is a two-dimensional convolutional layer, and the process is as follows:

[0043]

[0044] In the formula, and are the outputs of the main feature and the auxiliary feature after 2D CNN, and they will be used as the inputs of the time module. Both W(·) and b(·) are learnable parameters.

[0045] Furthermore, a parallel time convolutional structure is built to represent the time relationship between electricity consumption. By introducing a three-dimensional dynamic adjacency matrix, the complex spatial relationship between electricity consumption is represented, and a dynamic graph convolutional structure is constructed. During the construction of the parallel time convolutional structure, a dilated causal convolutional structure is adopted, and a gated DCCL formed by two parallel dilated causal convolutional layers (Dilation Causal Convolution Layer, DCCL) is controlled by a gating mechanism. Then, after being processed by two activation functions, Tanh(·) and σ(·), an element-wise product operation is finally performed. The whole process is as follows:

[0046]

[0047] In the formula, is the output of the main feature of the time convolutional layer of the time module at l th , and the output of the auxiliary feature of the time convolutional layer of the time module at l th is W f，l and W g，l are the learnable parameters of the convolutional filter, and σ(·) is the Sigmoid activation function.

[0048] By introducing a three-dimensional dynamic adjacency matrix, the complex spatial relationship between electricity consumption is represented, and a dynamic graph convolutional structure is constructed. This process considers a matrix factorization method - the tri-modal factorization method. The goal of this idea is to reconstruct the original data matrix X through the core matrix G and the factor loading matrices A, B, and C, and at the same time decompose it into low-dimensional representations of multiple modes. Its core formula is as follows:

[0049]

[0050] In the formula, x i，j，k is an element in the original data matrix X (three-dimensional), representing at sample i th , variable j th and kth Observed value at a time point, a i，m Is an element in the factor loading matrix A (two-dimensional), describing the relationship between pattern i (sample) and derived pattern m (factor), b j，p Is an element in the factor loading matrix B (two-dimensional), describing the relationship between pattern j (variable) and derived pattern p (factor), c k，q Is an element in the factor loading matrix C (two-dimensional), describing the relationship between pattern k (time point) and derived pattern q (factor), g m，p，q Is an element in the core matrix G (three-dimensional), describing the interaction relationship between factors. It is the core part of the model and captures the interaction effects between different patterns.

[0051] In the present invention, the reverse three-modal factor analysis method is used to reconstruct the original adjacency matrix by constructing 3 factor loading matrices and 1 core matrix Incorporate the time relationship into the construction of the adjacency matrix. The process specifically includes:

[0052] Step (1) Utilize the temporal correlation of residential electricity consumption in adjacent time periods to share the electricity consumption relationship of 24 hours in a graph structure, thereby dividing all time steps into cycles of 24 hours. Here, N t = 24, thus reducing the complexity of the original adjacency matrix to

[0053] Step (2) The goal of the three-modal factor analysis method used is to reconstruct the original data matrix A by constructing the core matrix G m，p，q and the factor loading matrices A t，m , B i，p , C j，q to achieve the purpose of dimensionality reduction. The core formula is as follows: t，i，j Wherein,

[0054]

[0055] In the formula, is the time slot factor loading matrix, is the input node factor loading matrix, is the output node factor loading matrix and is the core matrix, where N i is the number of input nodes, N o is the number of output nodes, d is the embedding dimension. Here, the factor loading matrices A t，m , B i，p , C j，q are two-dimensional matrices, and G m，p，q is a three-dimensional matrix,

[0056]

[0057] The above two formulas are respectively the non - negativity processing and row normalization processing of the constructed adjacency matrix, obtaining the final three - dimensional dynamic adjacency matrix for use.

[0058] Step (3) is to construct three three - dimensional adjacency matrices that meet the model's usage according to the three - dimensional adjacency matrix construction method in step (2). They are the three - dimensional adjacency matrices of the main feature and the auxiliary features and Here, since multiple auxiliary features are used, only one is used to represent all auxiliary features for convenience of description. And is the three - dimensional adjacency matrix used in the multi - feature fusion module. Different from the conventional method of directly splicing the auxiliary features onto the main feature for prediction, in the multi - feature fusion module, by constructing to dynamically simulate the spatio - temporal relationship between the main feature and the auxiliary features, N m represents the number of main feature nodes, and N c represents the number of auxiliary nodes.

[0059] Step (4) is to generate a graph structure based on the three three - dimensional adjacency matrices obtained in step (3). A graph g=(v, e) consists of a set of node sets v and a set of edge sets e. Among them, v i , v j ∈v represents a node, and ε ij ∈e represents the edge between v i and v j . And these three graph structures are exactly the connections between the electricity consumption of multi - household residences and their auxiliary features, representing their spatial dependence relationships. In this process, a diffusion graph convolutional network is used to handle the spatial dependence between residences by aggregating the hidden states of adjacent residences. The formula is:

[0060]

[0061] In the formula, is the input of the l th dynamic graph convolutional layer, and O l is the output updated by the dynamic graph convolutional layer. Similarly, the output of the main feature is The output of the auxiliary feature is W k is the learnable parameter for k th diffusion, and L is the total number of diffusion steps, thus generating the main feature graph g m , the auxiliary feature graph g c and the fused graph g cm of the main feature and the auxiliary feature.

[0062] Then, a multi-feature fusion module is designed. This module takes into account the spatio-temporal dependencies that also exist between the nodes of the auxiliary features. In each layer, a separate spatio-temporal block is assigned to the main feature and the auxiliary feature, and the outputs of the main feature and the auxiliary feature are obtained. These outputs are respectively input into their respective multi-feature fusion modules, namely g cm and g m , and after an aggregation process, the finally fused output is obtained. The formula is:

[0063]

[0064] In the formula, DGC(·) is the dynamic graph convolutional layer, and are the outputs of the main feature and the auxiliary feature after passing through the dynamic graph convolutional layer, is the output after multi-feature fusion. Agg(·) represents the aggregation process of the two results. The purpose of using direct addition is to make the calculation of the entire model concise and maintain a high representation ability. The adjacency matrix at time slot t is expressed as where ω(t) is a function to obtain the number of time slots in a period, W represents the learnable parameter, represents the output result after passing through the multi-feature fusion module.

[0065] Finally, the entire multi-feature fusion spatio-temporal graph neural network (DFF-GWN) is constructed. The data after passing through the spatio-temporal module is output as the prediction result through the activation function and the two-dimensional convolutional network. In order to improve the performance and training efficiency of the model in the entire network, residual connections and skip connection structures are also added. The skip connection structure integrates the spatio-temporal dependencies of different time scales. The formula is:

[0066]

[0067] In the formula, concat(·) is the concatenation operation, is the output result of the main feature of the time module in each layer, and L is the number of spatio-temporal layers. The finally obtained output result is as follows:

[0068]

[0069] In the formula, O m is the output of the model after passing through the skip connection, is the predicted value of the residential electricity consumption in the future Q time steps.

[0070] In the example of the present invention, common evaluation indicators in the neural network verification method are selected to verify the effectiveness of the present invention. The mean absolute percentage error (MAPE[%]) and the mean absolute error (MAE[kW h]) are respectively selected for verification. The formulas for both are:

[0071]

[0072] In the formula, is the predicted value of residential electricity consumption at time t, and y t is the actual value of residential electricity consumption at time t, and T is the number of time steps.

[0073] As shown in the prediction result Table 1, under the condition that other conditions of the model are the same, the model proposed in the present invention is compared with three advanced models, namely LSTM, CNN-GRU, and Ada-GWN, to predict the average electricity consumption of 15 households in the next 1 hour, and evaluation indicators are used to judge the effectiveness of the model.

[0074] Table 1 Results Table of Model Evaluation Indicators

[0075] Model MAPE [%] MAE [kW h] LSTM 8.68 0.78 CNN-GRU 9.86 0.89 Ada-GWN 7.62 0.72 The prediction model of the present invention 6.36 0.62

[0076] It can be seen from the results in Table 1 that the effect achieved by the prediction model GFF-GWN of the present invention is the best, which can prove the advantages of the present invention.

Claims

1. A method for predicting short-term residential electricity consumption based on feature fusion graph neural network, characterized in that: The method comprises the following operations: First, using the public data set provided by the OpenEI official website, we selected the historical electricity consumption data of 15 residences in Louisiana, USA, and two auxiliary feature historical data of the residences. The historical electricity consumption data of these 15 residences includes the historical electricity consumption data per hour in 2012. Each household records the electricity consumption of various electrical equipment per hour, with a total of 8760 data points recorded. The two auxiliary features are Electricity: Facility and Electricity: HVAC. The main feature is the total electricity consumption, which is the prediction object. Secondly, the residential electricity consumption dataset and the two auxiliary feature datasets are processed through the input module respectively and compiled into a set of datasets. The existing residential electricity consumption dataset and the two auxiliary feature datasets are and Collectively referred to as X 1:T , where 1:T represents the historical time step of T. In the following analysis, only To represent two auxiliary features, the input module is passed through a two-dimensional convolution layer. The process is as follows: In the formula, and are the outputs of the main features and auxiliary features after 2DCNN, which will be used as the input of the time module. W(·) and b(·) are both learnable parameters; Furthermore, a parallel time convolution structure is built to represent the temporal relationship between power consumption. In the process of building the parallel time convolution structure, an expanded causal convolution structure is adopted, and a gated expanded causal convolution layer formed by two parallel expanded causal convolution layers is controlled by a gating mechanism. Then, the gated expanded causal convolution layer is processed by two activation functions, Tanh (·) and σ (·), and finally an element-by-element product operation is performed. The whole process is as follows: In the formula, Yes th The output of the main features of the temporal convolutional layer of the temporal module, l th The output of the auxiliary features of the temporal convolutional layer of the temporal module is W f,l and W g,l is the learnable parameter of the convolution filter, σ(·) is the Sigmoid activation function; Using the inverse trimodal factor analysis method, the original adjacency matrix is ​​reconstructed by constructing three factor loading matrices and one core matrix. The process specifically includes: Step (1) utilizes the temporal correlation of residential electricity consumption in adjacent time periods and divides all time steps into 24-hour cycles, where N t =24, thus converting the original adjacency matrix The complexity is reduced to Step (2) uses the inverse trimodal factor analysis method to construct the core matrix G m,p,q And the factor loading matrix A t,m ,B i,p ,C j,q To reconstruct the original adjacency matrix The formula is as follows: In the formula, is the slot factor loading matrix, is the input node factor loading matrix, is the output node factor loading matrix, is the core matrix, where N i is the number of input nodes, N o is the number of output nodes, d is the embedding dimension, and the factor loading matrix A t,m ,B i,p ,C j,q is a two-dimensional matrix, G m，p，q is a three-dimensional matrix, The above two formulas are respectively the non-negative processing and row normalization processing of the created adjacency matrix, and the final three-dimensional adjacency matrix is ​​obtained; Step (3) is to construct three three-dimensional adjacency matrices that meet the needs of the model based on the three-dimensional adjacency matrix construction method of step (2), namely, the main feature three-dimensional adjacency matrix Auxiliary feature three-dimensional adjacency matrix The three-dimensional adjacency matrix used in the multi-feature fusion module In the multi-feature fusion module, by constructing To dynamically simulate the spatiotemporal relationship between the main feature and the auxiliary feature, N m Represents the number of main feature nodes, N c Represents the number of auxiliary feature nodes; Step (4) generates a graph structure based on the three three-dimensional adjacency matrices obtained in step (3), a graph It consists of a set of nodes v and a set of edges e, where v i , v j ∈v represents a node, ε ij ∈e represents v i and v j The edges between them are the connections between the electricity consumption of multiple residential buildings and their auxiliary features, and the diffusion graph convolutional network is used to handle the spatial dependency between residential buildings by aggregating the hidden states of adjacent residential buildings. The formula is: In the formula, Yes th The input of the dynamic graph convolution layer, O l is the output after being updated by the dynamic graph convolution layer, and the main feature output is The auxiliary feature output is W k k th The learnable parameters of diffusion, K is the total number of diffusion steps; Then, a multi-feature fusion module is designed. This module takes into account the spatiotemporal dependency between the nodes of the auxiliary features. After the aggregation process, the fused output is obtained. The formula is: Where DGC(·) is the dynamic graph convolution layer, is the output after multi-feature fusion, Agg(·) represents the aggregation process of two results, and the adjacency matrix of time slot t is expressed as where ω(t) is a function that captures the number of time slots in a cycle, W represents a learnable parameter, Represents the output result after the multi-feature fusion module; Finally, the entire multi-feature fusion spatiotemporal graph neural network is formed. In order to improve the performance and training efficiency of the model, residual connection and jump connection structures are added to the entire network. The jump connection structure integrates different time dependencies. The formula is: In the formula, concat(·) is the concatenation operation, is the output result of the main feature of the time module in each layer, L is the number of spatiotemporal layers, and the final output result is as follows: In the formula, O m is the output of the model after skip connection, It is the predicted value of residential electricity consumption in the future Q time steps; the model is verified using public data sets, and the effectiveness of the model is judged by evaluation indicators.

2. According to the method for predicting short-term residential electricity consumption using a feature fusion graph neural network according to claim 1, it is characterized in that: There are many auxiliary features in the data set used. For the selection of these auxiliary features, the correlation coefficient analysis method is used to calculate the coefficient between each auxiliary feature and the power consumption. The two auxiliary features selected are determined according to the size of the coefficient. The correlation coefficient comprehensively considers the Pearson coefficient and the Spearman coefficient. The formulas of the Pearson coefficient and the Spearman coefficient are as follows: In the formula, r and ρ are Pearson coefficient and Spearman coefficient respectively, x i and i are the observation values ​​of auxiliary features and electricity consumption, and are the means of auxiliary features and power consumption, d i is x i and i , T is the time step.

3. According to the method of claim 1, the feature fusion graph neural network short-term residential electricity consumption prediction method is characterized in that: The processing of short-term residential electricity consumption and its auxiliary characteristics is to normalize the data separately, then splice all the data in the feature dimension and compile them into a set of data sets for easy input into the model.

Images